1. No category

M PRA
Munich Personal RePEc Archive
A Bayesian Model of Sample Selection
with a Discrete Outcome Variable
Obrizan Maksym
Kyiv School of Economics, Kyiv Economics Institute
2010
Online at http://mpra.ub.uni-muenchen.de/28577/
MPRA Paper No. 28577, posted 4. February 2011 07:56 UTC
A Bayesian Model of Sample Selection
with a Discrete Outcome Variable
Maksym Obrizana,b , PhD
December 26, 2010
Affiliations:
a
b
Assistant Professor, Kyiv School of Economics (KSE)
Senior Economist, Kyiv Economics Institute (KEI)
Address:
13 Yakira St., Suite 318, Kyiv, 04119, Ukraine
Phone:
(+380) 44-492-8012
Fax:
(+380) 44-492-8011
E-mail:
mobrizan@kse.org.ua
1
Abstract
Relatively few published studies apply Heckman’s (1979) sample selection model to
the case of a discrete endogenous variable and those are limited to a single outcome
equation. However, there are potentially many applications for this model in health,
labor and financial economics. To fill in this theoretical gap, I extend the Bayesian
multivariate probit setup of Chib and Greenberg (1998) into a model of non-ignorable
selection that can handle multiple selection and discrete-continuous outcome equations.
The first extension of the multivariate probit model in Chib and Greenberg (1998)
allows some of the outcomes to be missing. In addition, I use Cholesky factorization of
the variance matrix to avoid the Metropolis-Hastings algorithm in the Gibbs sampler.
Finally, using artificial data I show that the model is capable of retrieving the parameters used in the data-generating process and also that the resulting Markov Chain
passes all standard convergence tests.
Keywords: Bayesian computing, latent variable models, Markov Chain Monte Carlo,
Bayesian modeling, sample selection model, multivariate probit
2
1
Introduction
The problem of sample selection applies whenever a dependent variable is missing as a result
of a non-experimental selection process. Economists have been aware for a long time that
estimating such a model by ordinary least squares leads to inconsistent estimates.1 Gronau
(1974) seems to be among the first to recognize this problem, but Heckman (1979) offers a
truly pioneering work with a simple two-step estimator that has been widely used for more
than three decades.
Empirical applications of the sample selection model, however, have been mostly limited
to the case of a continuous endogenous variable in an outcome equation. In addition, the
majority of papers deal with a single selection and a single outcome equation in the sample
selection model. In practice, sample may be chosen based on more than one criterion, or
more than one outcome equations may be considered.
In this paper I offer a Bayesian model of sample selection with two additional features.
First of all, it allows binary dependent variable in the outcome equation as well.2 Secondly,
adding extra selection or outcome equations with dichotomous or continuous dependent
variables is straightforward. This extension is crucial in the presence of multiple selection
equations, as explained below. These two extensions seem to be an important contribution
to the existing literature with potential applications in health, labor and related empirical
economic research.
While technical issues limited the use of sample selection models with multiple binary
dependent variables, their applicability is potentially very wide. Consider an example from
health economics where a researcher is interested in joint estimation of two or more binary
morbid health events (for example, hip fracture and stroke) in a sample of Medicare-eligible
older Americans. Suppose further that she observes those outcomes only for respondents
who allowed her to access their Medicare claims. Clearly, joint estimation of the two health
events (outcome variables of interest) with a third equation for being in the analytic sample
(selection equation) tends to be more efficient than estimating them equation-by-equation.3
More importantly, in order to obtain consistent estimates all of the selection equations have
1
2
There is no such problem if the disturbances in two equations have zero correlation.
I review some earlier work on a discrete outcome variable in Heckman’s (1979) model and explain how
my model differs further on.
3
This is a standard result in seemingly unrelated regression model, which does not apply if the explanatory
variables are the same or if the correlation/covariance terms are zero.
3
to be included.
Consider another example from financial economics. Suppose that a credit card company
studies the probability of default (outcome equation) for respondents who received a credit
card offer. The first selection equation may be if they accepted the offer and applied for a
card, and the second whether their application was approved by the bank. In this model,
the agent can default only if she was approved for a credit card, which in turn is possible
only if she has responded to such an offer.
In labor economics it might be of interest to study employment discrimination (observed
for candidates that seek a job) and wage discrimination (observed for candidates that seek a
job and are hired). These two outcome equations can be estimated together with selection
equation (if a candidate is seeking a job or not). All these and related models can be
estimated in the framework developed in this paper.
How is the problem of sample selection accounted for in the multivariate probit model?
To continue with the health economics example, suppose that there exists some unobserved
factor that affects both the probability of being selected into a sample and of having a morbid
health event. If healthier individuals are more likely to allow access to their Medicare claims,
say, because of a better cognitive function, then estimating the probability of a morbid health
event only for the observed subsample is not representative of the entire population, as only
its healthier part is considered.
From the discussion above, it is apparent that in order to consistently estimate a model
with sample selection, it is necessary to account for an omitted variable problem. In general,
the sample selection problem arises if the unobserved factors determining the inclusion in
the subsample are correlated with the unobservables that affect the endogenous variable of
primary interest (Vella 1998). In the current paper the specification error of omitted variable
resulting from selection is dealt with by considering the unobserved omitted variable as a
part of the disturbance term and then jointly estimating the system of equations accounting
for the correlations in the variance-covariance matrix.
The multivariate probit model can be used to handle multiple correlated dichotomous
variables along the lines of Ashford and Sowden (1970) and Amemiya (1974). It seems, however, that the potential of this model has not been fully realized despite its connection to
the normal distribution, which allows for a flexible correlation structure. As noticed in Chib
and Greenberg (1998), at least part of the problem in earlier applications arose from the
4
difficulties associated with evaluating the likelihood function by classical methods, except
under simplifying assumptions like equicorrelated responses, as in Ochi and Prentice (1984).
Chib and Greenberg (1998) describe how the model can be reformulated in a Bayesian
context using the technique of data augmentation (discussed in Albert and Chib [1993],
among others). The discrete dependent variable in the probit model can be viewed as the
outcome of an underlying linear regression with some latent dependent variable (i.e. unobserved by the researcher). Consider a decision to make a large purchase, as in Greene (2003,
p. 669). If the benefits outweigh the costs (benefits-costs>0) then the latent dependent
variable is positive and the purchase is made (the observed discrete outcome is 1), and vice
versa. If the researcher makes a further assumption that the disturbance term in the model
with the latent dependent variable has a standard normal distribution, then the univariate
probit model results. The extension to the multivariate case is relatively straightforward.
The latent variables are clearly not observed, but their distributions are specified to be
normal. Chib and Greenberg (1998) use this fact and re-introduce the latent variable back
into the multivariate probit model. In a typical Bayesian model the prior distribution of the
parameters and the likelihood function are used to obtain the joint posterior distribution,
which combines the information from the prior and the data. Chib and Greenberg (1998) find
the joint posterior distribution of the multivariate probit model as the product of the prior
distribution of the parameters and augmented likelihood function. The latter is obtained
as the product of normal distributions for latent variables taken over all respondents in
the sample. It is easy to show that, after integrating over the latent variables, the joint
posterior distribution of the parameters is exactly the same as the posterior distribution
obtained without introducing any latent variables (see Koop, Poirier and Tobias [2007] for
related examples). The computational advantage of this method — it does not require the
evaluation of the truncated multivariate normal density — is the greater the more discrete
dependent variables are included into the model.
Using the full conditional posterior distributions of the coefficient vector, along with
elements in the variance matrix and the latent data, it is possible to construct a Markov
Chain Monte Carlo (MCMC) algorithm and simulate the parameters jointly with the latent
data. In the Chib and Greenberg (1998) formulation, the conditional posterior distribution
for the elements in the variance matrix has a nonstandard form and the authors use a
Metropolis-Hastings algorithm to draw those elements. This paper modifies the Chib and
5
Greenberg (1998) procedure by using the Cholesky factorization of the variance matrix. This
allows a convenient multivariate normal representation of the parameters that are used to
obtain the variance matrix, which considerably facilitates estimation.
Another complication in the sample selection model follows from the fact that some of
the dependent binary variables in the outcome equation are not observed given the selection
rule into the sample. The posterior distribution of the latent data can be used to simulate those missing observations conditional on the covariance structure of the disturbance
term. Consider first an individual t with complete data in m × 1 vector of binary responses
y.t = (y1t , ..., ymt )0 for all selection and outcome equations. The Chib and Greenberg (1998)
procedure implies that at each MCMC simulation the latent vector ye.t = (e
y1t , ..., yemt )0 is
drawn from the truncated multivariate normal distribution with a m × 1 mean vector and
m × m covariance matrix Σ.4 The distribution is truncated for the ith element yeit to (−∞, 0]
if the binary outcome yit = −1 and to (0, +∞) if yit = 1. Now suppose that individual t
has missing binary outcome yit for some i. The only difference with the case of an observed
binary outcome yit comes from the fact that the conditional multivariate normal distribution
for yeit is no longer truncated in the ith dimension. That is, if yit is missing for some i, then
the latent variable yeit is unrestricted and can take any value in the interval (−∞, ∞).
Identification of the parameters is an important issue in models of discrete choice. It is
well-known that the multivairate probit model is not likelihood-identified with unrestricted
covariance matrix. Even though the formulation of the variance matrix in this paper uses
only m(m − 1)/2 identified parameters, this turns out not to be sufficient for identification.
Meng and Schmidt (1985) offer an elegant treatment of the problem of identification in the
censored bivariate probit model using the general principle in Rothenberg (1971) that the
parameters in the model are (locally) identified if and only if the information matrix is
nonsingular. The conclusion in Meng and Schmidt (1985), that the bivariate probit model
with sample selection is in general identified, applies also with my parameterization of the
model.
This paper is organized as follows. Section 2 reviews the literature on sample selection
especially on extensions to models with discrete outcome equation and Bayesian treatment.
Section 3 sets up the model and derives the details of the multivariate probit estimator.
4
The mean vector for individual t is a product of m × k matrix of covariates and a k × 1 vector of
coefficients to be defined later.
6
Section 4 develops the Gibbs sampler. Section 5 considers the problem of identification
in greater detail. Finally, section 6 provides an illustrative example and the last section
concludes the discussion.
2
Discrete Outcome Variable in Sample Selection Model
2.1
Classical treatment of a discrete outcome variable
The model of incidental truncation, which is another name for sample selection model, has
been widely used in economic applications when the variable of interest is observed only
for people who are selected into a sample based on some threshold rule. Heckman’s (1979)
treatment of sample selection model is also a standard topic in most modern econometric
textbooks (such as Greene 2003 and Wooldridge 2002).
However, there are relatively few applications of Heckman’s (1979) model to discrete (and
count) data and Greene (2008) reviews a handful of such models, starting with Wynand and
van Praag (1981). In a recent application to teen employment, Mohanty (2002) uses the
formulation in Meng and Schmidt (1985), which is very similar to the bivariate probit model
with sample selection in Wynand and Praag (1981). In Mohanty (2002) the applicant i
for a job can be selected (SELi = 1) or not (SELi = 0) only if she has applied for a job
(SEEKi = 1). Both discrete variables are modeled as the latent variables y1i (SEEKi = 1
if y1i > 0 and SEEKi = 0 otherwise) and y2i (SELi = 1 if y2i > 0 and SELi = 0 otherwise)
that have bivariate normal distribution with correlation coefficient ρ.
Estimating the hiring equation (SELi ) only for the subsample of teens who applied for
a job (SEEKi = 1) produces inconsistent estimates as long as ρ 6= 0. Indeed, univariate
probit shows misleading evidence of employment discrimination against Black teens, which
disappears when participation and hiring equations are estimated jointly (Mohanty 2002).
Another relevant example in classical econometrics is Greene (1992), who refers to an
earlier paper by Boyes, Hoffman and Low (1989). The (part of the) model in Greene (1992)
is bivariate probit where the decision to default or not on a credit card is observed only for
cardholders (and not the applicants that were rejected by a credit card company).
Terza (1998) is another important reference in this literature. He develops a model for
count data that includes an endogenous treatment variable. For example, the number of
trips by a family (count variable of interest) may depend on the dummy for car ownership
7
(potentially endogenous). In this case the dependent variable for car ownership in the first
equation appears as explanatory variable in the equation for the number of trips and the
two equations are estimated jointly. Terza (1998) compares three estimators for this model:
full information ML, non-linear weighted least squares (NWLS) and a two-stage method of
moments (TSM) similar to Heckman’s (1979) estimator.5
The setup in Terza (1998) can be potentially used in models of discrete choice with
sample selection, as in a recent paper by Kenkel and Terza (2001). Kenkel and Terza (2001)
use a two-step estimator in the model of alcohol consumption (number of drinks) with an
endogenous dummy for advice (from a physician to reduce alcohol consumption). The first
stage is univariate probit for receiving advice and the second stage applies non-linear least
squares to the demand for alcohol (number of drinks). Kenkel and Terza (2001) find that
advice reduces alcohol consumption in the sample of males with hypertension, and the failure
to account for the endogeneity of advice would mask this result.
Munkin and Trivedi (2003) discuss the problems with different estimators of selection
models with discrete outcome equation. The first class of models, which uses moment-based
procedures, results in inefficient estimates and does not allow the estimation of the full set
of parameters in the presence of correlated multiple outcomes. A second possibility is a
weighted nonlinear instrumental variable approach that has not been very successful because of difficulties in consistent estimation of weights (Munkin and Trivedi 2003). Finally,
simulated maximum likelihood method requires a sufficient number of simulations for consistency where it is not clear what is “...the operational meaning of sufficient” (Munkin and
Trivedi 2003, p. 198).
It seems that none of the models discussed so far allows multiple correlated discrete
dependent variables in the presence of sample selection (except for the bivariate case). The
approach that I adopt in this paper is to extend the Bayesian multivariate probit model in
Chib and Greenberg (1998), allowing for some missing responses. I review existing Bayesian
treatments of sample selection in the next subsection and then provide further details on
Chib and Greenberg (1998).
5
The estimators are listed in the order of decreasing efficiency and computational difficulty. NWLS
estimator may result in correlation coefficient being greater than one in absolute value.
8
2.2
Bayesian treatment of Heckman model
Recent Bayesian treatments of sample selection model are almost exclusively based on
Markov Chain Monte Carlo (MCMC) methods with data augmentation.6 The idea of data
augmentation was introduced by Tanner and Wong (1987), and used in Bayesian discrete
choice models starting (at least) from Albert and Chib (1993).7 Latent variables in these
models are treated as additional parameters and are sampled from the joint posterior distribution. In these models, however, the joint posterior distribution for parameters and
latent variables typically does not have a recognizable form. Gibbs sampler is an MCMC
method used when the joint posterior distribution can be represented as a full set of (simpler)
conditional distributions. It is possible then to obtain the sample from the joint posterior distribution by iteratively drawing from each conditional distribution, given the values obtained
from the remaining distributions. The model developed herein shares the two aforementioned
features (data augmentation and Gibbs sampling) and simultaneous equation structure with
previous studies by Li (1998), Huang (2001) and van Hasselt (2008).
Li (1998) develops Bayesian inference in the following simultaneous equation model with
limited dependent variables (SLDV):
y1∗ = y2 γ1 + X1 δ1 + u1
y2∗ = X2 δ2 + u2 ,
(1)
where y1∗ is of Tobit type (a researcher observes y1 = y1∗ if y1∗ > 0 and y1 = 0 otherwise) and
y2∗ is of probit type (the researcher observes y2 = 1 if y2∗ > 0 and y2 = 0 otherwise).8 The
vector of disturbances (u1 , u2 )0 is assumed to follow bivariate normal distribution with the
variance of u2 set to 1 for model identification:
Ã
!
2
σ11
σ12
Σ=
,
σ12 1
6
Earlier developments in Bayesian statistics model selection by means of various weight functions. For
example, Bayarri and DeGroot (1987 and four other papers, as cited in Lee and Berger 2001) mostly concentrate on indicator weight function: potential observation is selected into a sample if it exceeds a certain
threshold. Bayarri and Berger (1998) develop nonparametric classes of weight functions that are bounded
above and below by two weight functions. Lee and Berger (2001) use the Dirichlet process as a prior on the
weight function.
7
Notice that the selection equation in a Heckman-type model is univariate probit.
8
To avoid the confusion with my parameters later on, I use different Greek letters from those used in the
original papers throughout the literature review.
9
2
where σ11
is the variance of u1 and σ12 is the covariance between u1 and u2 . Decomposing
the joint bivariate distribution of (u1 , u2 )0 into the product of the marginal distribution of u2
and the conditional distribution of u1 |u2 allows convenient blocking in the Gibbs sampler.
This decomposition in Li (1998), together with the more convenient reparametrization of the
variance matrix
Ã
Σ=
2
σ 2 + σ12
σ12
σ12
1
!
,
appear repeatedly in later studies. With these changes the model is now re-defined as
y1∗ = y2 γ1 + X1 δ1 + u2 σ12 + v1
y2∗ = X2 δ2 + u2 ,
(2)
2
2
with u2 = y2∗ − X2 δ2 , σ 2 = σ11
− σ12
, and v1 ∼ N (0, σ 2 ). In the resulting Gibbs sampler
with data augmentation, all conditional distributions have recognizable forms that are easy
to draw from (multivariate normal, univariate truncated normal and gamma).9
Huang (2001) develops Bayesian seemingly unrelated regression (SUR) model, where
dependent variables are of the Tobit type (researcher observes yij = yij∗ if yij∗ > 0 and
yij = 0 otherwise). The Gibbs sampler with data augmentation in Huang (2001) consists of
multivariate normal, Wishart and truncated multivariate normal distributions.
In the paper by van Hasselt (2008), two sample selection models — with unidentified parameters and with identified parameters — are compared.10 The idea behind the first model
is borrowed from McCulloch and Rossi (1994), who used a similar approach in multinomial
probit context. The output from the Gibbs sampler is used to approximate the posterior
distribution of the identified parameters. The model with identified parameters in van Hasselt (2008) uses marginal-conditional decomposition of the disturbance terms together with
9
Chakravarti and Li (2003) apply this model to estimate dual trade informativeness in futures markets.
Probit equation estimates a trader’s decision to trade on her own account and tobit equation measures
her (abnormal) profit from her own account trading. Chakravarti and Li (2003) did not find significant
correlation between a dual trader’s private information and her abnormal profit.
10
Another interesting paper by van Hasselt (2005) compares the performance of sample selection and twopart models (when two equations are estimated independently) in a Bayesian setup. In classical econometrics
Leung and Yu (1996) provide conclusive evidence against negative results in Manning, Duan and Rogers
(1987) who claim that two-part model performs better than sample selection model even when the latter
is the true model. Leung and Yu (1996) show that problems with sample selection model are caused by a
critical problem in the design of experiments in Manning, Duan and Rogers (1987).
10
more convenient parameterization of the variance matrix, as in Li (1998).11 The major contribution of van Hasselt (2008) is relaxing the normal distribution assumption in the sample
selection model via mixture of normal distributions. I do not follow that route and my model
remains fully parametric.
In all the papers cited above the outcome variable is continuous and not discrete. There
are two Bayesian papers with discrete outcome variable (and multiple outcome equations)
that are worth mentioning: Munkin and Trivedi (2003) and Preget and Waelbroeck (2006).
Munkin and Trivedi (2003) develop a three-equation model with the first equation for
count data (the number of doctor visits), the second equation for a continuous variable (the
associated health expenditures) and the third equation for a dummy variable (the type of
health insurance plan). The selection problem — demand for health care that potentially
depends on the type of health insurance — is modeled by using an (endogenous) dummy
variable for private health plan. There is no problem of missing dependent variable for respondents that are not in the sample (i.e. who did not purchase private insurance). Neither
of the correlation coefficients for private health plan with two variables of interest is statistically different from zero and the type of insurance does not affect the level of health care
use (Munkin and Trivedi 2003).12
Preget and Waelbroeck (2006) develop a three-equation model with application to timber
auctions. There are two binary dependent variables (if a lot received any bids and, conditional
on receiving at least one bid, if a lot received two or more bids) and one continuous variable
(highest bid for a lot) with an endogenous dummy variable for the number of bids. Preget
and Waelbroeck (2006) comment that in such models the likelihood function is not always
well behaved, especially in the direction of the correlation coefficients.13 While in Preget and
11
McCulloch, Polson and Rossi (2000) show that fully identified multinomial probit model comes at a cost:
higher autocorrelation in the Markov Chain.
12
In a later work, Deb, Munkin and Trivedi (2006), perhaps dissatisfied with a sample selection model,
use a two-part model with endogeneity in a similar context.
13
Consider the following sequential probit model: the second binary outcome is missing for all respondents
whose first outcome is “No.” The third binary outcome, if present, is missing for all respondents who answered
“No” in the second equation and so on. Waelbroeck (2005) argues that in this model the likelihood function is
not globally concave and flat in some directions, which limits practical applicability of the model. Notice that
in a two-equation case, sequential probit is the same model as censored probit except that the two models
may have different interpretation. Keane (1992) discusses similar computational issues in multinomial probit
model.
11
Waelbroeck (2006) the correlation coefficients are never statistically different from zero, they
find that their Bayesian algorithm “...yields a remarkably stable coefficient for the binary
endogenous variable and was able to deal with irregularities in the likelihood function.”
Two conclusions seem to follow from my review of relevant studies. First of all, there exist
serious computational difficulties when the sample selection model with multiple dichotomous dependent variables is estimated by methods of classical econometrics. For example,
Munkin and Trivedi (2003) comment on difficulties associated with estimating their model in
a simulated maximum likelihood framework. This provides strong motivation for a Bayesian
econometric methodology and also explains why models similar to mine are typically estimated in a Bayesian and not classical tradition. Second, even in the Bayesian literature,
there seem to be no published papers that can be used directly to estimate a model with
three or more dichotomous dependent variables. This constitutes an important contribution
of the current paper.
While my work shares the methods with previous studies (data augmentation, Gibbs
sampling and simultaneous equation structure) it comes from a different area — multivariate
probit model developed in Chib and Greenberg (1998). The next section introduces the
multivariate probit in Chib and Greenberg (1998) and provides the extensions that make it
applicable in the sample selection model.
3
Multivariate Probit and Sample Selection
Suppose that a researcher observes a set of potentially correlated binary events i = 1, ..., m
over an independent sample of t = 1, ..., T respondents. Consider the multivariate probit
model reformulated in terms of latent variables as in Chib and Greenberg (1998). For each
of the events i = 1, ..., m define a T × 1 vector of latent variables yei. = (e
yi1 , ..., yeiT )0 and a
T × ki matrix of explanatory variables Zi where each row t represents a 1 × ki vector Zit .
Then each latent variable can be modeled as
yei. = Zi βi + εi. ,
(3)
where εi. is a vector of disturbance terms that have normal distribution. There is potential
correlation in the disturbance terms for respondent t across events i = 1, ..., m coming from
some unobserved factor that simultaneously affects selection and outcome variables. Let
12
ye.t = (e
y1t , ..., yemt )0 be the vector of latent variables for respondent t such that
ye.t ∼ Nm (Zt β, Σ),
(4)
where Zt = diag(Z1t , ..., Zmt ) is an m × k covariate matrix, βi ² Rki is an unknown parameter
P
0 0
vector in equation i = 1, ..., m with β = (β10 , ..., βm
) ² Rk and k = m
i=1 ki , and Σ is the
variance matrix.
The sign of yeit for each dependent variable i = 1, ..., m uniquely determines the observed
binary outcome yit :
yit = I(e
yit > 0) − I(e
yit <= 0) (i = 1, ..., m),
(5)
where I(A) is the indicator function of an event A. Suppose it is of interest to evaluate the
probability of observing a vector of binary responses Y. = (Y1 , ..., Ym )0 for indivial t. Chib
and Greenberg (1998) show that the probability y.t = Y.t can be expressed as
Z
Z
...
φm (ye.t |Zt β, Σ)dye.t ,
Bmt
(6)
B1t
where Bit ² (0, ∞) if yit = 1 and Bit ² (−∞, 0] if yit = −1. Define Bt = B1t × ... × Bmt .
Alternatively, the probability y.t = Y.t can be expressed without introducing latent variables as
Z
Z
pr(y.t = Y.t |β, Σ) =
...
Amt
φm (w|0, Σ)dw,
(7)
A1t
where φm (w|0, Σ) is the density of a m-variate normal distribution and Ait is the interval
defined as
(
Ait =
(−∞, Zit βi )
if yit = 1,
(Zit βi , ∞)
if yit = −1.
The multidimensional integral over the normal distribution in (7) is hard to evaluate by
conventional methods.14
Instead of evaluating this integral, Chib and Greenberg (1998) use the formulation in (6)
and simulate the latent variable ye.t from the conditional posterior distribution with mean
Zt β and variance matrix Σ. This distribution is truncated for the ith element to (−∞, 0] if
14
Quadrature method is an example of nonsimulation procedure that can be used to approximate the
integral. Quadrature operates effectively only when the dimension of integral is small, typically not more
than four or five (Train 2003). The GHK simulator is the most widely used simulation method after Geweke
(1989), Hajivassiliou (as reported in Hajivassiliou and McFadden 1998) and Keane (1994).
13
the observed outcome is yit = −1 and to (0, +∞) if yit = 1. The current model also assumes
that yit = 0 when the response for event i is missing for t.
It is important to understand what missing binary response means in terms of the latent
data representation. If respondent t has missing binary response yit for some i then no
restriction can be imposed on the latent normal distribution in the ith dimension. Then
the vector ye.t is simulated from the m-variate normal distribution with the same mean and
variance as in the complete data case but the distribution is not truncated for the ith element.
For the case of missing outcome i the latent variable yeit can take any value in the interval
(−∞, ∞).15
The multivariate model of incidental truncation can not be estimated using only the
observed data because the endogenous selection variables are constant and equal to 1. Now,
due to simulated missing data one can estimate the variance matrix Σ, which is the focus
of the procedure to account for sample selection. The covariances in Σ effectively adjust for
sample selectivity in the outcome equations by controlling for unobserved heterogeneity.
The issue of sample selection arises whenever the unobserved factors determining the
inclusion in the sample are correlated with the unobservables that affect the outcome variable(s) of primary interest (Vella 1998). The critical idea in the current work is to account
for selection in binary outcome equation(s) by jointly estimating selection and outcome
equations while controlling for possible unobserved effect through multivariate probit with
correlated responses. If the covariance terms belong to the highest posterior density region,
this indicates the presence of unobserved effect and, hence, sample selection bias.
The elements in the variance matrix in the Chib and Greenberg (1998) formulation do
not have the conditional posterior distribution of a recognizable form, which forces them
to employ a Metropolis-Hastings algorithm. This paper makes the technical advance that
allows convenient multivariate normal representation of the parameters used to obtain the
variance matrix. Consider the Cholesky factorization of the inverse of the variance matrix
Σ−1 = F˘ · F˘ 0 where F˘ is the lower triangular matrix. If the diagonal elements of F˘ are
arrayed in a diagonal matrix Q then Σ−1 = F˘ Q−1 Q2 Q−1 F˘ 0 = F Q2 F (Greene 2003). In the
current work the variance matrix is defined by F which is a lower triangular matrix that has
15
This methodology allows for continuous endogenous variables as well. In this case yejt is trivially set to
the observed yjt for a continuous variable j in each iteration of the MCMC algorithm introduced below.
14
ones on the main diagonal and D−1 = Q2 which is a diagonal matrix. Then
Σ = (F 0 )−1 DF −1 ,
(8)
with D = diag{d11 , ..., dmm } and F is lower triangular


1
0
0 ··· 0


 f21
1
0 ··· 0 




1 ··· 0 .
F =  f31 f32
 .
..
..
. . . .. 
 ..
.
.
. 


fm1 fm2 fm3 · · · 1
Finally, consider the system of m equations





ye = 
|{z} 

T m×1
ye1.

ye2. 

,
.. 
. 

yem.




Z =
|{z}

T m×k

Z1
0
···
0
0
..
.
Z2 · · ·
.. . .
.
.
0
..
.
0
0
···



,








β =
|{z} 

k×1
Zm
β1
β2
..
.



,


βm
so that the model can be represented as
ye = Zβ + ε,
where k =
Pm
i=1
ki and




ε =
|{z}

T m×1

(9)

ε1
ε2
..
.



.


εm
Under the maintained assumption of the normally distributed vector ε it follows that
ε|(β, F, D, Z) ∼ N (0, (F 0 )−1 DF −1 ⊗ IT ).
4
(10)
Deriving the Gibbs Sampler
Consider a sample of m × T observations y = (y.1 , ..., y.T ) that are independent over t =
1, ..., T respondents but are potentially correlated over i = 1, ..., m events. Given a prior
density p(β, F, D) on the parameters β, F and D the posterior density is equal to
p(β, F, D|y) ∝ p(β, F, D)p(y|β, Σ),
15
(11)
where p(y|β, Σ) =
QT
t=1
p(y.t |β, Σ) is the likelihood function. Define y.t = (yst , yot ), where
yst and yot are selection and outcome variables with some of the y.t ’s missing. In this
representation the evaluation of the likelihood function is computationally intensive from
a classical perspective. Albert and Chib (1993) developed an alternative Bayesian framework that focuses on the joint posterior distribution of the parameters and the latent data
p(β, F, D, ye1 , ..., yeT |y). It follows then that
p(β, F, D, ye|y) ∝ p(β, F, D)p(e
y |β, Σ)p(y|e
y , β, Σ)
(12)
= p(β, F, D)p(e
y |β, Σ)p(y|e
y ).
It is possible now to implement a sampling approach and construct a Markov chain from the
distributions [e
y.t |y.t , β, Σ] (t ≤ T ), [β|y, ye, Σ] and [F, D|y, ye, β].
With unrestricted F or D matrix the multivariate probit model is not identified. The
observed outcomes y.t for respondent t depend only on signs but not magnitudes of the latent
data ye.t . In a multivariate probit model with m equations only m(m − 1)/2 parameters
in the variance matrix are identified. Consider the following transformation of the model
F 0 ye.t ∼ N (F 0 Zt β, D), where D is some unrestricted diagonal matrix. The latent regression
has the form F 0 ye.t = F 0 Zt β + D1/2 ε.t , where ε.t is m-variate normal with a zero mean vector
and an m × m identity variance matrix. However, pre-multiplying this equation by α > 0
results in αF 0 ye.t = F 0 Zt (αβ) + αD1/2 ε.t which is the same model corresponding to the same
observed data y.t . Since the parameters in D1/2 cannot be identified, D is set to identity
matrix extending the logic from the univariate probit model in Greene (2003).16
The posterior density kernel is the product of the priors and the augmented likelihood
in equation (12).17 The parameters in β and F are specified to be independent in the prior.
Let the prior distribution for β be normal φk (β|β, B −1 ) with the location vector β and the
precision matrix B.
It is convenient to concatenate the vectors below the main diagonal in F matrix as


F
 2:m,1 
 F3:m,2 


Fvector = 
,
.
..




Fm,m−1
16
Observe that this is not sufficient for identification and later I give an example from Meng and Schmidt
(1985) when the model is not identified with two equations.
17
The term “augmented likelihood” emphasizes the fact that the likelihood includes latent variables.
16
where Fi+1:m,i for i = 1, ..., m − 1 represents elements from i + 1 to m in column i. The prior
¡
¢
distribution of Fvector is assumed to be m(m−1)
-variate normal
2
Fvector ∼ N (F vector , H −1 ).
(13)
In this expression F vector is the prior mean of the normal distribution, and the prior variance
matrix H −1 is block-diagonal with

H
0
···
0
 2:m,2:m

0
H 3:m,3:m · · ·
0

H=
.
.
..
..
..
..

.
.

0
0
· · · H 1,1




.


This precision matrix has (m − 1) × (m − 1) matrix H 2:m,2:m in the upper left corner and the
matrix dimension is decreasing by one in each consequent block on the main diagonal. The
lower right matrix H 1,1 is a scalar. The posterior density kernel is now
o
1
− (β − β)0 B(β − β)
2
n 1
o
· |H|1/2 exp − (Fvector − F vector )0 H(Fvector − F vector )
2
T
o
n
Y
1
y.t − Zt β)0 Σ−1 (e
y.t − Zt β) I(e
y.t ² Bt ).
· |Σ|−T /2
exp − (e
2
t=1
|B|1/2 exp
n
(14)
A Gibbs sampler is constructed by drawing from the following conditional posterior distributions: the vector of coefficients β, the Fvector from the variance matrix decomposition
and the latent vector ye.t for each respondent t ≤ T .18
In a typical iteration the Gibbs sampler initiates by drawing the vector of the coefficients
β conditional on Fvector and ye.t obtained from the previous draw. The posterior distribution
of β comes from the posterior density kernel and is normal
−1
β|(e
y , Σ) ∼ Nk (β|β, B ),
where B = B +
PT
t=1
−1
Zt0 Σ−1 Zt and β = B (Bβ +
PT
t=1
Zt0 Σ−1 ye.t ). In this last expression it
is understood that for each t , ye.t ² Bt .
18
Technical Appendix A.1 provides complete details of the Gibbs sampler derivation.
17
(15)
To obtain the conditional posterior distribution of F , an alternative expression for the
density of ye is useful:
p(e
y |y, β, F, D) ∝ |Σ|
−T /2
T
Y
n
exp
t=1
−1
0 T /2
= |F D F |
T
Y
o
1
0 −1
− (e
y.t − Zt β) Σ (e
y.t − Zt β) I(e
y.t ² Bt )
2
n
exp
t=1
=
=
T Y
m
Y
t=1 i=1
m
Y
n
exp
n
exp
i=1
o
1
− ε0t F D−1 F 0 εt I(e
y.t ² Bt )
2
o
1
0
− (εt,i + Fi+1:m,i
εt,i+1:m )2
2
o
1X
0
−
(εt,i + Fi+1:m,i
εt,i+1:m )2 ,
2 t=1
T
(16)
where for each t , ye.t ² Bt . In this derivation the restriction D = Im is already imposed.
Then the posterior conditional distribution of Fvector is also normal
Fvector |(y, ye, β) ∼ N¡ m(m−1) ¢ (F vector , H
−1
).
(17)
2
The conditional posterior normal distribution has the posterior precision matrix
 P

T
0
ε
ε
0
·
·
·
0
 t=1 t,2:m t,2:m PT

0


0
ε
ε
·
·
·
0


t=1 t,3:m t,3:m
H =H +
.
..
..
..
...


.
.
.


PT
0
0
0
···
t=1 εt,m εt,m
The posterior mean of the normal distribution is equal to
 P
F vector = H
−1
H F vector − H
−1
T
t=1 εt,2:m εt,1
PT
t=1 εt,3:m εt,2





 P
T
..
.




.


t=1 εt,m εt,m−1
Finally, the latent data ye.t are drawn independently for each respondent t ≤ T from the
truncated multivariate normal distribution as described in Geweke (1991). The algorithm
makes draws conditional on Zt , β and F as well as ye.t obtained in the previous draw. The
multivariate normal distribution is truncated to the region defined by the m × 2 matrix [a, b]
with a typical row i equal to (0, ∞), if yit = 1 and (−∞, 0) if yit = −1. If yit is not observed,
then row i is (−∞, ∞).
18
Thus, this work extends Chib and Greenberg (1998) in the following two ways: (i) it
permits missing outcome variables ye.t , and (ii) it re-parameterizes the variance matrix in
terms of more convenient multivariate normal Fvector that is used to obtain Σ.
5
The Problem of Identification
Identification is an important issue in models of discrete choice. Meng and Schmidt (1985)
in their elegant article offer an excellent treatment of identification in a bivariate probit
model under various levels of observability. Meng and Schmidt (1985) rely on the general
principle in Rothenberg (1971) that the parameters are (locally) identified if and only if the
information matrix is nonsingular. In particular, their Case Three: Censored Probit is very
similar to the following bivariate sample selection model: the binary variable of interest y2t
is observed for respondent t only if she is selected in the sample (y1t = 1).19
Let F t = F (Z1t β1 , Z2t β2 ; f21 ) specify the bivariate normal cumulative distribution function and Φ(Zht βh ) specify the univariate standard normal cumulative distribution function
with h = 1, 2 for respondent t. Recall that the sign of yeit perfectly predicts yit and one can
write
p(y|e
y) =
T
Y
I(e
y1t > 0)I(e
y2t > 0)I(y1t = 1)I(y2t = 1)
t=1
+I(e
y1t > 0)I(e
y2t ≤ 0)I(y1t = 1)I(y2t = −1) + I(e
y1t ≤ 0)I(y1t = −1).
The likelihood function in the bivariate model can be obtained after I integrate over ye in the
19
I employ different parametrization of the variance matrix and, thus, the parameters have to be scaled
to be comparable with Meng and Schmidt (1985).
19
following way
Z
Z
Z
p(y, ye|β, Σ)de
y=
B
=
T Z
Y
∞
Z
∞
p(y|e
y )p(e
y |β, Σ)de
y
B
h
I(e
y1t > 0)I(e
y2t > 0)I(y1t = 1)I(y2t = 1)
−∞
t=1
p(y|e
y , β, Σ)p(e
y |β, Σ)de
y=
B
−∞
i
+I(e
y1t > 0)I(e
y2t ≤ 0)I(y1t = 1)I(y2t = −1) + I(e
y1t ≤ 0)I(y1t = −1)
·f (Z1t β1 , Z2t β2 ; f21 )de
y1t de
y2t
Z
Z
T
Y ∞ ∞
=
I(y1t = 1)I(y2t = 1)f2 de
y1t de
y2t
+
t=1 0
T Z 0
Y
t=1
+
=
−∞
T Z
Y
t=1
T
Y
0
ÃZ
!
∞
I(y1t = 1)I(y2t = −1)f2 de
y1t de
y2t
0
0
I(y1t = −1)φ(Z1t β1 )de
y1t
−∞
F (Z1t β1 , Z2t β2 ; f21 )I(y1t =1)I(y2t =1)
t=1
·[Φ(Z1t β1 ) − F (Z1t β1 , Z2t β2 ; f21 )]I(y1t =1)I(y2t =−1)
·[1 − Φ(Z1t β1 )]I(y1t =−1) ,
where f = f (Z1t β1 , Z2t β2 ; f21 ) is bivariate normal density function and φ(·) is univariate
normal density function. Define qit =
yit +1
2
for i = 1, 2 and take the natural logarithm of
this expression to obtain
ln L(Z1 β1 , Z2 β2 ; f21 ) =
(18)
T
i
Xh
q1t q2t ln F t + q1t (1 − q2t ) ln[Φ(Z1t β1 ) − F t ] + (1 − q1t ) ln[1 − Φ(Z1t β1 )] ,
t=1
which is equivalent to equation (6) in Meng and Schmidt (1985) except for the correlation
coefficient ρ being replaced by the parameter f21 that enters F t as defined below equation
(8). The general conclusion reached by Meng and Schmidt (1985) is that the parameters
in this model are identified except in certain “perverse” cases. First of all, peculiar configurations of the explanatory variables may cause nonidentification, but this problem can be
addressed only given the data at hand. Second, nonidentification may be caused by certain
combinations of parameters in the model. For example, the censored bivariate probit model
20
with my parametrization is not identified when Z1t β1 = √−f21 2 Z2t β2 for all respondents t
and I show this result in Technical Appendix A.2.
20
1+f21
The information matrix is then singular
because the row for the second intercept (i.e. for (k1 + 1)th term) is the last row (i.e. for the
parameter f21 ), divided by a constant. In this particular example the problem of nonidentification does not arise as long as the set of explanatory variables is not the same in two
equations.
Meng and Schmidt (1985) comment that there might also be other combinations of parameters or particular configurations of explanatory variables leading to nonidentification.
Since it is not possible to foresee all such problems a priori, it is the responsibility of the
researcher to check if the parameters in the model are identified. However, it is very reassuring that the sample selection model is generally identified, except in some (not very likely)
cases.
6
Experiments with Artificial Data
The purpose of the experiment with artificial data is to study if the model can retrieve the
parameters and the correlation coefficient that are used to generate the data when some of
the outcome variables are missing. It is also of interest to assess the convergence properties
of the model. I construct the following bivariate probit model with sample selection. Let
y2t be the dichotomous dependent variable of interest that is observed only if the selection
variable y1t is equal to 1.
For this experiment I generate t = 1, ..., 500 independent latent variables (e
y1t , ye2t )0 from
p
2
the bivariate normal distribution with mean µt = [Z1t β1,. 1 + f21
, Z2t β2,. ]0 , where a 1 × 3
vector Zit contains intercept, one discrete and one continuous variable as described below
and βi,. = [βi,1 βi,2 βi,3 ]0 for i = 1, 2. Each equation contains the intercept denoted βi,1 ,
continuous variable βi,2 and discrete variable βi,3 . Continuous variable in each equation is
drawn from the normal distribution with µ = −0.5 and σ = 2. Discrete variable takes values
of −1 and 1 with equal probability. All continuous and discrete variables are independent
from each other. The coefficients used to generate the artificial data are provided in the
20
Another example of nonidentification given in Meng and Schmidt (1985) is when there are only intercepts
included in all equations. While such a model cannot be used in a meaninful way for economic analysis, it
provides an interesting limiting case when all the covariate coefficients go to zero.
21
second column of Table 1. The correlation coefficient is set to 0.5 with the corresponding
value of f21 ≈ −0.5774. Finally, the 2 × 2 covariance matrix is the same for all respondents
and is set to
"
Σ=
2
1 + f21
−f21
−f21
1
#
.
p
2
Observe that the true parameters of the first equation are multiplied by 1 + f21
and in
p
2
each simulation I normalize the draws of β1,. by 1 + f21
obtained in the same draw. After I
obtain the 500 × 2 matrix of the latent variables ye, I convert it into the matrix of “observed”
dichotomous dependent variables y which is used in the simulator. The coefficients that
were chosen place approximately one third in each of the three bins (yes, yes), (yes, no) and
(no, missing).
The implementation of the Gibbs sampler is programmed in the Matlab environment with
some loops written in C language. All the codes successfully passed the joint distribution
tests in Geweke (2004). The results in this section are based on 24,000 draws from the
posterior (the first 6,000 draws were discarded as burn-in iterations). The prior for i = 1, 2
vector of coefficients βi,. is mutlivariate normal with the mean vector set to zeros and the
variance matrix equal to the identity matrix of dimension 3. The prior for f21 is standard
normal distribution.
The results of the experiment are shown in Figures 1-3 and Table 1.21 The simulator
works quite well in this experiment with low autocorrelation and stable results with histograms centered almost at the values of the parameters used to generate the data. Geweke’s
convergence diagnostic test (Geweke 1992) does not indicate problems with the convergence
of the Markov Chain. The only slight problem is that the mean of the correlation coefficient
ρ in the sample obtained from the joint posterior distribution (0.23) is somewhat lower than
the value of 0.5 used to obtain the artificial data but it still belongs to the 95% highest
posterior density interval.
7
Concluding Remarks
This paper develops a sample selection model for discrete or mixed continuous-discrete outcomes with multiple outcome and selection equations. To facilitate the estimation of a
21
To obtain some of the statistics I used the MATLAB program momentg.m by James LeSage.
22
resulting multivariate probit model, a Bayesian reformulation in terms of latent variables
is extended from the Chib and Greenberg (1998) paper that offers a convenient simulation
procedure aimed at resolving the problems of evaluating the integral of multivariate normal
density. The essence of the method is to jointly simulate the parameters and the latent variables from conditional posterior distributions using a Markov Chain Monte Carlo algorithm.
If there is any unobserved heterogeneity for each agent t, it is properly accounted for as a
part of the disturbance terms by the covariance structure of the variance matrix resulting
from a joint estimation of a system of equations.
This paper also makes two technical advances to the Chib and Greenberg (1998) setup by
(i) adding some missing binary responses and (ii) simplifying the estimation of the variance
matrix via a multivariate normal representation of the elements in the lower triangular matrix
from the Cholesky factorization of Σ−1 . I also discuss how the results on identification in
Meng and Schmidt (1985) apply in the bivariate probit model with sample selection.
In addition to introducing the multivariate probit model with sample selection, this paper
also offers some interesting topics for further research. In particular, it might be of interest
to further study the identification in the case of three and more equations, which clearly
depends on the selection rule into a sample. The likelihood is different in each particular
case and extensive study of this topic along the lines of Meng and Schmidt (1985) may be
rewarding. Alternatively, some of the potentially interesting topics in empirical health and
labor economics outlined in the introduction can be done with little (or no) modification of
the model in this paper.
Acknowledgements
I would like to thank Professor Fredric Wolinsky for his help and financial support through
NIH grant R01 AG-022913 in the preceding two years. This paper benefited from insightful comments of seminar participants at the Department of Health Management and Policy
at the University of Iowa (Iowa City, USA) in October of 2009 and Kyiv School of Economics/Kyiv Economics Institute (Kyiv, Ukraine) in March of 2010. John Geweke suggested
the extension of the Chib and Greenberg (1998) paper, provided help on some derivations
and also financial support in the Spring of 2009. Michelle Nourski assisted in editing the
manuscript. Any remaining errors are mine.
23
References
Albert, J., Chib S., 1993. Bayesian analysis of binary and polychotomous response data.
Journal of the American Statistical Association 88 (422), 669-679.
Amemiya, T., 1974. Bivariate probit analysis: Minimum chi-square methods. Journal of the
American Statistical Association 69 (348), 940-944.
Ashford, J., Sowden R., 1970. Multi-variate probit analysis. Biometrics (26), 535-546.
Bayarri, M., Berger J., 1998. Robust Bayesian analysis of selection models. Annals of Statistics (26), 645-659.
Bayarri, M., DeGroot M., 1987. Bayesian analysis of selection models. The Statistician (36),
137-146.
Boyes, W., Hoffman, D., Low S., 1989. An econometric analysis of the bank credit scoring
problem. Journal of Econometrics (40), 3-14.
Chakravarty, S., Li K., 2003. A Bayesian analysis of dual trader informativeness in future
markets. Journal of Empirical Finance 10 (3), 355-371.
Chib, S., Greenberg E., 1998. Analysis of multivariate probit models. Biometrika 85 (2),
347-361.
Deb, P., K. Munkin, Trivedi P., 2006. Bayesian analysis of the two-part model with endogeneity: Application to health care expenditure. Journal of Applied Econometrics (21),
1081-1099.
Geweke, J., 1989. Bayesian inference in econometric models using Monte Carlo integration.
Econometrica (57), 13171339.
Geweke, J., 1991. Efficient simulation from the multivariate normal and Student-t distributions subject to linear constraints, in: Keramidas, E. (Ed.), Computing Science and
Statistics: Proceedings of the 23rd Symposium on the Interface. Fairfax: Interface Foundation of North America, Inc.
Geweke, J., 1992. Evaluating the accuracy of sampling-based approaches to the calculation
of the posterior moments, in: Berger, J., Dawid J., Smith A. (Eds.), Bayesian Statistics 4.
Oxford: Clarendon Press.
Geweke, J., 2004. Getting it right: Joint distribution tests of posterior simulators. Journal
24
of the American Statistical Association (99), 799-804.
Greene, W., 1992. A statistical model for credit scoring. WP No. EC-92-29, Department of
Economics, Stern School of Business, New York University.
Greene, W., 2003. Econometric Analysis, fifth ed. Prentice Hall, Upper Saddle River.
Greene, W., 2008. Econometric Analysis, sixth ed. Pearson Education, Upper Saddle River.
Gronau, R., 1974. Wage comparisons - a selectivity bias. The Journal of Political Economy
82 (6), 1119-1143.
Hajivassiliou, V., McFadden D., 1998. The method of simulated scores for the estimation of
LDV models. Econometrica (66), 863-896.
Heckman, J., 1979. Sample selection bias as a specification error. Econometrica 47 (1),
153-161.
Huang, H.-C., 2001. Bayesian analysis of the SUR tobit model. Applied Economics Letters
(8), 617-622.
Keane, M., 1992. A note on identification in the multinomial probit model. Journal of
Business and Economic Statistics 10 (2), 193-200.
Keane, M., 1994. A computationally practical simulation estimator for panel data. Econometrica (62), 95-116.
Kenkel, D., Terza, J., 2001. The effects of physician advice on alcohol consumption: Count
regression with an endogenous treatment effect. Journal of Applied Econometrics 16 (2),
165-184.
Koop, G., Poirier, D., Tobias J., 2007. Bayesian Econometric Methods, first ed. Cambridge
University Press: New York.
Lee, J., Berger, J., 2001. Semiparametric Bayesian analysis of selection models. Journal of
the American Statistical Association (96), 1269-1276.
Leung, S., Yu, S., 1996. On the choice between sample selection and two-part models. Journal of Econometrics (72), 197-229.
Li, K., 1998. Bayesian inference in a simultaneous equation model with limited dependent
variables. Journal of Econometrics (85), 387-400.
Manning, W., Duan, N., Rogers W., 1987. Monte Carlo evidence on the choice between
sample selection and two-part models. Journal of Econometrics (35), 59-82.
McCulloch, R., Polson, N., Rossi P., 2000. A Bayesian analysis of the multinomial probit
model with fully identified parameters. Journal of Econometrics (99), 173-193.
25
McCulloch, R., Rossi P., 1994. An exact likelihood analysis of the multinomial probit model.
Journal of Econometrics (64), 207-240.
Meng, C., Schmidt P., 1985. On the cost of partial observability in the bivariate probit
model. International Economic Review (26), 71-86.
Mohanty, M., 2002. A bivariate probit approach to the determination of employment: a
study of teen employment differentials in Los Angeles county. Applied Economics 34 (2),
143-156.
Munkin, M., Trivedi P., 2003. Bayesian analysis of a self-selection model with multiple
outcomes using simulation-based estimation: An application to the demand for healthcare.
Journal of Econometrics (114), 197-220.
Ochi, Y., Prentice R., 1984. Likelihood inference in a correlated probit regression model.
Biometrika 71 (3), 531-543.
Preget, R., Waelbroeck P., 2006. Sample selection with binary endogenous variable: A
Bayesian analysis of participation to timber auctions. Working Paper ESS-06-08, Telecom
Paris.
Rothenberg, T., 1971. Identification in parametric models. Econometrica (39), 577-591.
Tanner, M., Wong W., 1987. The calculation of posterior distribution by data augmentation.
Journal of American Statistical Association (82), 528-540.
Terza, J., 1998. Estimating count data models with endogenous switching: Sample selection
and endogenous treatment effects. Journal of Econometrics 84 (1), 129-154.
Train, K., 2003. Discrete Choice Methods with Simulation, first ed. Cambridge University
Press: New York.
van Hasselt, M., 2005. Bayesian sampling algorithms for the sample selection and two- part
models. Working paper, Department of Economics, Brown University.
van Hasselt, M., 2008. Bayesian inference in a sample selection model. Working paper,
Department of Economics, The University of Western Ontario.
Vella, F., 1998. Estimating models with sample selection bias: A survey. The Journal of
Human Resources 33 (1), 127-169.
Waelbroeck, P., 2005. Computational issues in the sequential probit model: A Monte Carlo
study. Computational Economics 26 (2), 141-161.
Wooldridge, J., 2002. Econometric Analysis of Cross Section and Panel Data, first ed. The
MIT Press: Cambridge.
26
Wynand, P., Praag B.V., 1981. The demand for deductibles in private health insurance.
Journal of Econometrics (17), 229-252.
27